I had the Need for Speed, and boy have I been having some fun today.
I've just received my new (just off cutting-edge) Dell Precision T5500 workstation, and have moved all my drives over to it. This dual hex core 3.47 GHz Xeon X5690 system upgrades my previous dual quad core Dell Precision T5400 3.16 GHz Xeon X5460 setup.
The T5400 was no slouch, but benchmarks like Passmark showed the T5500 to be up to twice as powerful overall the T5400, so when I saw a price I liked I took the plunge.
Unfortunately, the VisionTek ATI Radeon HD 7850 I had ordered along with it was defective out of the box, so for now I'm still using my prior 5670 card. But without a video card upgrade, this gives me the unique opportunity to test just how Photoshop and other operations respond to the increased processing power as well as RAM, using the same disks, same software installation, same video card and drivers. So I set out to determine just how much this upgrade would affect the Photoshop work I do, which I'm sure will be interesting to folks here...
Here's what I have found:
A few things truly do take only half as long, though most are taking more like 2/3 as long and a few are surprisingly at about the same speed (probably implying my video card is the bottleneck). All times are T5500 vs. T5400:
T5500 (this was a good stress test, the system sounded like it was ready to take off):
Interestingly, disk access to my already hyper-fast SSD array has actually increased a good bit, implying the disks and controller were capable of delivering and accepting data faster than the T5400 system itself. As with many things listed above, the new system is about 1.5x to 2x faster on small operations. For example, if I select all the entries in the root of drive C:, then choose Properties, the T5400 would enumerate all 511,000 files and stop counting up in about 32 seconds, at a rate of about 16,000 files / second. The new system finishes that count in 20.4 seconds, at a rate of 25,000 files / second. This means that my already very fast file access just feels snappier for everything. My software builds now take only about 60% as long as they did before.
Looks like another week or more before I get my much faster video card, at which time I'll run through it all again.
I haven't tried stitching a panorama yet. Right now as I write this it's doing the render I listed above. The pano is next. I think the big RAM is going to help a lot with that.
Any other things you'd like me to try? I won't be able to compare directly until I get the older system back together (been spending all my time playing with the new rig).
Well, I didn't invent the 3D rendering facility of Photoshop, nor have I set the rendering target to be reasonable. For pretty much anything practical it looks good enough to eat in about 45 seconds, at which point you can just click on something to stop the rendering.
The point of the exercise wasn't how quickly or poorly it rendered this particular object but the comparison between the two machines.
Yeah, the panos are of interest! Try just building some reallllly big ones! I've just finished assembling, masking, adjusting and tiling out ortho images in the 15 - 30 GB ea. range using my single i7 quad core 5500 (SSD boot 10K data & scratch drives) with 24GB ram. Being an hourly employee made the wait times tolerable. I was system monitoring all along and have noticed very little cpu usage vs ram getting maxed out (and multiple 70GB temp files!). I kinda knew that when I spec'ed this box that I wanted more ram than cpu, but what you have now is what I want in a year. The 12 cores would come in very handy for large image processing in Lightroom too..
A Photomerge stitch of 25 Canon raw files at 6144 x 4096 pixels each and 16 bits/channel yielding a 344 megapixel output finished in 6 minutes and 46 seconds (down from over 10 minutes on my older system). It used 39 GB of RAM and still used the same amount of scratch space (100 GB+) . I was a bit disappointed that given the huge increase in RAM the result was only incrementally faster.
The system was still perfectly responsive for doing other things while it was cranking away on this Photomerge, implying Photoshop is not using all the resources it can.
More comparitive test results... Conversion of blocks of raw files through Camera Raw this time.
Conversion of 50 Canon raw files with Camera Raw 7.2 RC, saved directly into 6144 x 4096 x 16 bits/channel PSD files (148 megabytes each).
64 bit ACR 7.2RC: 141.6 seconds (2.83 seconds per conversion)
32 bit ACR 7.2RC: 229.4 seconds (4.59 seconds per conversion)
64 bit ACR 7.2RC: 215.6 seconds (4.31 seconds per conversion)
32 bit ACR 7.2RC: 345.6 seconds (6.91 seconds per conversion)
Notably the conversion used only about 5% of the system's I/O speed capability, and about 25% of the processing capacity, implying Camera Raw not doing anywhere near the multithreading it could be.
Do you have Lightroom? I'd be real curious if someone with a box like this could see how long it would take to import and 1/1 thumbnail 100 or so 300MB image files at a crack (or ten and extrapolate the time). A pending job that will require the fastest possible visual analysis, correction and post processing of 1000s of this kind of aerial images seems to beg for the features available in LR but on my system its not very quick. I've heard the "PS likes ram and LR likes cpu's" so put 2+2 together and? Oh, and about GPU help?
Windows 7 x64 Ultimate. Same exact install I had before. After moving the drives over and booting up the new machine off them, Windows installed a bunch of drivers and rebooted, then I had to call Microsoft to re-activate it (they allowed 3 days, but I got it done right away). No problems whatsoever.
I'm not going to be moving to Windows 8 any time soon.
Hmmm, I actually expected you to move on. When I ran the Win8 trial, it didn't take long for me to uninstall it. There seemed little to offer (I ran CS6 trial on it s well) and the sudden move by MS to junk the years of GUI design and interaction in favor of touch , for systems that didn't have touch as the process to manipulate movement between apps, did nothing to convince me to go there. In fact, it convinced me to abandon the trial. I uninstalled it and wiped the drive and put XP back (because of legacy HW for which no updated drivers were offered by Epson).
But as a designer of apps for Windows, I rather would imagine you to, while finding some of Win 8 not to your liking, would nonetheless have to embrace it to continue offering apps.
Maybe you are thinking of retiring?
I think you've assumed a few things wrongly, Lawrence...
I may have mentioned at some time that I run Windows 8 in VMware virtual machines, so I actually DO run it, and I am fully equipped to develop for it. I'm just not embracing it for integrating my own work just yet (I'm waiting for the dust to settle a bit). VMware is a great "have cake and eat it too" application!
One of Windows 8's only attractions - that of being incrementally quicker to do many things than Windows 7, has been eclipsed for me by getting this new machine. It's utterly instantaneous to do everything, so I don't crave an improvement via software.
Just to get experience with how it runs on real hardware, I've already installed Windows 8 on my prior machine (T5400). Indeed, it's quicker to do many things than Windows 7 on the same machine, but even then it's not as quick as this 12 core goodness.
The flatten time: Is that for a 1.7G file? How many layers and what kind?
Relatively speaking, my ol' slowpoke here opens PSCS6 hot in 5-6 sec, so I want to compare some real stuff!
Another test I would love to do involves batch processing in DxO, but you don't have it.
I don't suppose you would run the trial version and do a batch test, would you? My average batch runs approx 9 sec for only camera corrections. With OpenCL turned off, it's closer to 20 seconds.
A user posted a test file over on the Photoshop.com forum with many layers for the purposes of stressing Photoshop to show the layer bounds bug in Photoshop CS6. Took me some looking, but I found the link:
Let me know how your system does with the tests I mentioned.
Regarding DXO, I'll keep your suggestion in mind, though I pretty aggressively avoid installing things I don't actually intend to use.
My replacement 7850 video card is due to come in tomorrow.
Open 31.4 sec
Save 37 sec
Flatten 15.8 sec
I saved the unflattened version after making a change in hue from 0 to1 so that I was actually saving an adjusted file.
The Open took at first, 46 sec, Starting PS first then opening the file. Subsequent openings with PS running dropped to 31 sec. Flatten stayed very close as well as Save.
Scratch generated 10G file
Opening is comparable to your earlier machine, (the 31 sec reopen) faster than your old but overall obviously either work station is faster, but not by a whole lot. Remember, I've using an Athlon IIx4, 2.88G cpu, 12G Ram, HDD drive. No hyperthreading.
A lot of this stuff is limited by RAM speed (the warm start reopens, for example, are likely completely cached in RAM). I'm still limited to 1333 MHz DDR3, 3 channels doubled by pairing, but it only can transfer so much data so fast. What speed are you running your RAM?
Even though I can sustain over one and a half gigabytes of data transfer to the SSD array per second, I've observed that a single thread can only read up to about 450 to 500 megabytes/second, so getting data in and out of a processor is challenging (and to be fair there are still some cache settings I need to fool with in the BIOS). This is why the system never seems to bog down - there's always another core and some reserve data transfer capacity, with essentially latency. It's a multitasker's dream, but not everything can work with multiple threads.
Adobe: Have you ever considered multi-threading your file I/O? Sounds horrendous with traditional storage, but with modern machines and SSDs it could kick your file access performance way up. It's a thought.
I went back to run other tests especially the brush test. I was somewhat limited by my own stroke speed but when I mastered that, I was consistently under 1 sec (best time 0.8sec) on a flattened version and about 1 sec on the top layer (where I could see what I was doing).
Open times varied but the most consistent was opening from Bridge. 45 sec. If I opened from Explorer there was a considerable lag time even showing PS and the timing was calculated from the moment PS showed up with the loader timing bar running.
Finally My system is actually faster than your old one on flattening, by a fair amount! So I would be shooting myself in the foot opting for a work station running DDR2.
My RAM setup is set at stock. I did run tests optimizing ram speed vs CL at Command Rate 1T and did see improvements but also instability. So I am at 1333MHz, CL9, 2T
So everything is stock.
Gives food for thought about any computer upgrades I was planning on giving this machine to my SO and building an Intel/Ivy Bridge for myself. The cost just for a processor and mobo will run several hundreds over an AMD Trinity/board combo, and Trinity is delivering pretty good performance for the PS filter, as a price point well below Intel (Amd's GPU on Trinity runs better than Intel)
All that changes if I were to decide I want to learn and run Premier.
I'm looking now to install a bigger C drive so I'll be looking at SSD's. Samsung seems to hold a strong edge.
OCZ brand SSDs are excellent. That's personal experience talking. Watch for a sale on 128GB Vertex 4 drives. The best I've seen them is $79.99 each. You could put, say, 4 of them in a system and have a lightning fast 512 GB C: drive from which you could run everything.
Got my 7850 video card today. Thank goodness this one works great. I had to add a $25 mini DisplayPort to DVI adapter to use my second (old) monitor, but now I swear my already pixel-crisp display is even sharper. Maybe that's just my rose colored glasses at work.
The 7850 is not night and day different from my 5670, but incrementally faster at everything. Everything on the desktop was instantaneous already, but the new card scores about 150% of what the older card did in 2D benchmarks, and more than 200% for 3D and OpenGL acceleration.
Photoshop consistently seems to come up a fraction of a second faster - perhaps that's setup time in the faster video card being saved.... I'm measuring 2.0 seconds now. I guess it'll have to be the next generation before I break into the 1.x range.
I do see a somewhat more interactive display in Photoshop's 3D, and I'll be darned but it seems like I'm seeing something I didn't see before - Image Based Light source reflections are now visible even in the draft rendering while moving things around. I don't recall seeing those before - maybe that's what the extra 1 GB of DDR5 on the video card gets you. It definitely helps with 3D operations not to have to do a little rendering to see how the reflections are going to look. Also, switching between line items in the 3D panel is now taking only a small fraction of a second.
Here are comparative benchmarks from my 4 most recent workstations (note that it's the same SSD array and blu-ray drive on both the T5400 and T5500).
Puget Systems just ran a test of video cards on a set of effects which employ the Mercury engine:
The 7750 and 7870 are FAPP tied (.75 sec difference )
The upgrade 13.0.1 makes a significant increase in performance.
I'll not be using the internal graphics of the Intel chip!
Message was edited by: Hudechrome
I almost got the 7750 as you have, except that I sense an advantage for the future in the extra 1 GB of DDR5. Things are only going to get more GPU-dependent - there's a WORLD of power in these GPUs.
I've been thinking about other, more decisive tests of your workstation vs conventionals. Premiere comes to mind, with a specific video to render.
I'm also looking at new computer build using ECC memory. They seem to be found in workstation environments and the mobos over $200.
An anecdote about ECC: Back early this year I began to have a system halt every week or so. The T5400 BIOS would report "Uncorrectable RAM Error in DIMM 5 or 6". A couple of times, on rebooting because of a Windows Update, it would report "RAM Errors in DIMM 5 or 6 were corrected".
After this happened a few times I replaced DIMMs 5 and 6 and had no more problems thereafter. It cost a whole $32 to ensure my system is reliable and trouble-free. That's no more than the price of a stupid cable nowadays, but the key point is that I was able to know just what was wrong..
Now imagine that happening with non-ECC RAM in a system not capable of detecting and reporting such a failure. Your application stores a 23 in a location and later reads back 2746237234 and your system just goes ahead and uses that corrupted data. It's almost a wonder anyone ever gets anything done with a non-ECC system.
RAM errors DO happen. I'll never use a computer without ECC again.
Actually, current research I did today on RAM indicate that the real difference between ECC and non versions is that ECC will detect and repair an error and non ECC will detect and report but not repair.
Perhaps this is an upgrade found within DDR3.
I'm looking into this and will grab links as I do it.
I've not found anything that hints that non-ECC RAM will do detection. The way that works the two go hand in hand (look up Hamming code), and it requires more hardware (read extra cost), so I doubt any non-ECC RAM does error detection. But by all means please point me to something that says otherwise.
If such a thing DOES exist, given the extra cost, you can be sure the feature will be clearly stated in the sales materials. If you don't have "parity" or "error detecting" somewhere on the package, I bet it ain't.
Keep in mind also that we're talking about PCs rather than just RAM in general across the world and across history. Keeping in mind the BIOS has to support the indication that there's been an error, as far as I know you have two choices: Non-ECC RAM and ECC RAM.
Here's snip of the averaged results:
Message was edited by: Hudechrome
Puget kindly published their benchmark action, after I asked. It's here:
It expects to work on the image found in this package:
Interestingly, near as I can tell when run as part of this action, the various new blur filters don't tax the GPU much at all. During the execution of the first 3 phases of the test my GPU activity went up to at most 5% only. Oil Paint and Lighting Effects caused blips of activity up to 20% - 25%. What's up with this? The implication is that Photoshop is still doing the lion's share of the work in the CPU.
Since the Puget benchmark results seem to measure individual core speed in addition to GPU performance, this makes comparisons between GPUs on different systems pretty much invalid. My system didn't score anything like the 7750 / 7870 results they published.
Did you check the website for the exact setup? (Silly question but I thought to ask anyway!)
I'll run the tests once a get past a networking problem between Win 7, XP and a new office "All-in-one" printer.
You just use the .atn file from the puget site.
Yes, I read how they set things up. Of course, I don't have the motherboards they used From what I could detect, it looks like the coordination of operations for each of these filters is done in the CPU in one thread (I could be wrong, but I didn't see that much CPU use overall, implying only one or a small number of cores was actually busy).
The eye-opener was that even when running these benchmarks the GPU is barely used. This is both bad news and good news. The bad news is that Photoshop is clearly not as fast as it could be. The good news is that there's so much room for improvement that we can expect good things for the speed of future versions of Photoshop.
I'm guessing you won't sense much of a difference with any modern GPU. It's not like waiting, say, 7 seconds for a filter to complete on a large image is particularly intrusive vs. 4 seconds.
The CPU individual core speed seems to matter more than anything else in these benchmarks.